PhD in digital design for efficient embedded machine learning processors - Belgium
Specifically, the project is currently looking for a PhDor post-doc to on resource-efficient digitalimplementations of machine learning processors, focused around Bayesian machinelearning (and more general: Probablistic Graphical Models),deep learning, and reinforcement learning.
Recently deep neural networks, such asconvolutional neural networks (CNNs) or long short-term memory (LSTM) networkshave gained enormous popularity in the signal processing community. In themicro-elecronics research domain this has sprouted attention on customizedprocessors for efficient embedded deep neural network inference. Our team haspublished several of these state-of-the-art processors over the past few years.
For the higher cognitive layers, whereoften sensor fusion takes place, a second machine learning paradigm isattractive: Bayesian learning and Probablistic Graphical Models. Thesetechniques enable to more smoothly inject expert knowedge into the system, andreason about the sensed information. White such “white box” classifiers areattractive from a knowledge point of view compared to the “black box” deepneural networks, their execution is still very computationally intensive ontraditional processors. And so far, no customized processors have been buildfor these workloads.
With this project, we want to enablethe power of Probablistic Graphical Models to embedded devices. This throughcustom processor design and hardwae accelerator design for both online learningand inference tasks. This research will hece require a combination ofalgorithmic innovations (dealing with reinforcement learning and ProbablisticGraphical Models) and hardware innovations (processor design, low-poweroptimization and chip tape-out).
We have already prooven in the field of Deep LearningProcessors, that such hardware/software co-optimization allows to save ordersof magnitude on energy efficiency. With this project, we want to achievesimilar gains for the next emerging deep learning technology beyond CNNs, DNNs,and RNNs, and enable the power of ProbablisticGraphical Models on embedded devices. In this project, we closelycollaborate with researchers from KU Leuven’s machine learning group DTAI, aswell as with UCLA’s machine learning group.
- Candidates must holda Masters degree inElectrical Engineering(or equivalent), witha background indigital design.
- Additional research/educationalexperience in computerarchitectures, machine learning or chip tape out is a strong plus.
- We are looking for a team player with the capability towork in an international research team.
- Excellent proficiencyin the English languageis also required, as well as good communicationskills, both oral and written.
· An exciting research environment,working on the intersection between emerging research domains (machine learning;processor design; ultra-low power chip and system design)
· A Ph.D. titlefrom a highly-rankeduniversity (after approximately4 years of successfulresearch)
· A thoroughscientific education,the possibility tobecome a world-classresearcher
· A KU Leuvenaffiliation, one ofthe largest researchuniversities of Europe
· The possibilityto participate ininternational conferencesand collaborations